Concentration of Power

The fear has two layers. The first is ordinary market concentration: the money, chips, data and specialist staff needed to train and serve a frontier model sit with a small number of companies and, at the state level, mainly with the United States and China. The second layer is what makes the first uncomfortable. If AI systems come to do the cognitive work that armies, bureaucracies and firms currently have to buy from people, then whoever controls those systems needs the cooperation of far fewer humans to get things done, and the ordinary levers of accountability lose their bite.

Tom Davidson, Lukas Finnveden and Rose Hadshar of Forethought set out the sharpest version in April 2025. They name three routes by which a small group could take durable control: systems built to be loyal to one person or a handful of people rather than to an institution; systems with hidden loyalties that pass inspection and act later; and exclusive access to capabilities in weapons development, strategic planning and cyber operations. Their claim is that such a seizure could in principle be carried out by a very small group, including inside an established democracy.^[1]

A second version needs no plot at all. Jan Kulveit, Raymond Douglas, Nora Ammann, Deger Turan, David Krueger and David Duvenaud argued in a 2025 paper, presented as a position paper at the International Conference on Machine Learning, that incremental capability gains alone could erode human influence over the economy, culture and the state. Their mechanism is substitution: as machine alternatives become competitive across labour, decision making and creative work, the explicit levers of human control such as voting and consumer choice weaken, and so do the implicit incentives that made institutions serve people in the first place.^[2]

The concern is not confined to critics of the industry. Dario Amodei, chief executive of Anthropic, published an essay in January 2026 that lists concentration of power alongside misuse and misalignment as one of the main risks of powerful AI, and explicitly includes his own company among the entities that could pose that risk. He treats domestic mass surveillance by AI as a line that democracies should regard as illegitimate.^[3]

The input side is the best measured part of the argument. Stanford's 2026 AI Index recorded global corporate AI investment of 581.7 billion dollars in 2025, with private investment growing 127.5 percent to 344.7 billion dollars. The United States accounted for 285.9 billion dollars of private investment, about 23 times the 12.4 billion dollars recorded for China. The same report notes that Chinese government guidance funds deployed an estimated 184 billion dollars into AI firms between 2000 and 2023, so the private figures understate Chinese spending.^[4]

In January 2025 the United States Federal Trade Commission published a staff report on the partnerships between Alphabet, Amazon and Microsoft on one side and Anthropic and OpenAI on the other. It described the consultation and exclusivity rights that cloud providers obtained, and flagged the risk that these arrangements shape access to computing resources and engineering talent while raising the cost to a developer of switching provider. The study was conducted under Section 6(b) of the FTC Act. The Commission did not allege illegality; it published the structure and named the competitive questions.^[6]

Cost is the mechanism that keeps the field small. Epoch AI's tracking of large training runs puts spending growth at about 2.4 times a year, which works out as the cost of the largest runs doubling roughly every eight months, a rate that puts the top tier out of reach of universities and most governments. Girish Sastry, Lennart Heim, Haydn Belfield and sixteen co-authors made the related governance point in a 2024 paper: computing power is detectable, excludable and quantifiable, and is produced through an extremely concentrated supply chain, which is why states and firms have started using it as a control point, and why control of it is itself a concentration risk.^[7][9]

The political channel is newer and easier to see. Public Citizen reported on 30 June 2026 that Leading the Future, the main AI industry super PAC in the United States midterm cycle, had raised 75.1 million dollars, including 50 million dollars from the venture firm Andreessen Horowitz and 25 million dollars from OpenAI president Greg Brockman and Anna Brockman. Counting the personal giving of Marc Andreessen and Ben Horowitz alongside the firm's, Public Citizen put the Andreessen Horowitz total across committees at 115.5 million dollars. The group describes its purpose as identifying and growing pro-AI candidates.^[8]

Government has moved in the same direction from the other side. On 14 July 2025 the Pentagon's Chief Digital and Artificial Intelligence Office awarded contracts with a ceiling of 200 million dollars each to Anthropic, Google, OpenAI and xAI. On 11 December 2025 the White House issued Executive Order 14365, Ensuring a National Policy Framework for Artificial Intelligence, directing the Attorney General to set up an AI Litigation Task Force within 30 days to challenge state AI laws, requiring the Secretary of Commerce to make states with onerous AI laws ineligible for parts of broadband funding, and instructing agencies to assess whether they can condition discretionary grants on states not enacting conflicting rules.^[10][11]

The Forethought mechanisms are worth taking seriously because they do not require any new physics, only reliable automation of tasks that currently need loyal people. A military that has replaced human crews with autonomous systems has replaced the veto that soldiers exercise when they refuse an unlawful order. A civil service that has replaced case officers with models has replaced the whistleblower. Davidson and colleagues note that hidden loyalties are hard to detect because a model can pass the tests it is given and behave differently in deployment, and because behaviour can propagate from one model generation to the next.^[1]

A slower path runs through the ordinary machinery of policy. If the industry with the deepest pockets in a given election cycle also supplies the tools that legislators, regulators and courts use to do their work, the normal remedies weaken from both ends. The December 2025 preemption order is a concrete instance: it does not ban state AI laws, but it creates a litigation unit whose purpose is to challenge them and links federal money to state restraint, which reduces the number of jurisdictions in which an alternative rule can be tried.^[11]

The labour channel matters because the threat to withdraw effort has historically been how ordinary people bargained with the powerful. Kulveit and colleagues argue that this bargaining position is not replaced by anything when machine substitutes are competitive, and that the loss shows up across the economy, culture and the state at once, with the effects in each domain reinforcing the others. The result they describe is not a coup but a drift, in which no single decision transfers power and no single decision can return it.^[2]

Chokepoints cut in more than one direction. The United States rescinded the Biden administration's AI diffusion rule in May 2025, announced on 8 December 2025 a one year waiver permitting exports of Nvidia H200 chips to China, and on 13 January 2026 the Bureau of Industry and Security released a licensing policy putting H200 exports to China on case by case review. Whichever way the policy runs, it shows that a small number of officials and one dominant chip designer can decide which countries get to build advanced systems at all.^[12]

The strongest objection is that capability is not staying scarce. The venture investor Tomasz Tunguz, writing in July 2026 and reading Chatbot Arena ratings from 2023 onward, describes open weight models repeatedly closing on the closed frontier before a new closed release pulls ahead again. He counts three large open weight releases in a single week of July 2026: Moonshot's Kimi K3 at 2.8 trillion parameters, Alibaba's Qwen 3.8 at 2.4 trillion, and Thinking Machines' Inkling, a 975 billion parameter multimodal model under an Apache 2.0 licence. On his pricing comparison the median open weight frontier model runs about 15 percent cheaper than GPT-5.2 and DeepSeek V4 Flash roughly 90 percent cheaper. If a near-frontier model can be downloaded and run cheaply, the bottleneck theory of power loses much of its force.^[13]

A second objection attacks the size of the prize. Daron Acemoglu, who shared the 2024 Nobel memorial prize in economics, has argued that the aggregate economic effect of current AI is modest. He told MIT Technology Review in February 2025 that he expected AI to raise United States gross domestic product by 1.1 to 1.6 percent over ten years, with an annual productivity gain of roughly 0.05 percent, and told Fortune in June 2026 that he still expected about 0.55 percent in total factor productivity gains over the decade. If the technology generates a small rather than transformative surplus, the wealth available to be concentrated is correspondingly small, and the political story rests on an economic premise that has not yet appeared in the national accounts. Acemoglu is not sanguine about AI and democracy, but his scepticism about the economics blunts the strongest versions of the lock-in argument.^[14][15]

Third, diffusion has been fast and wide rather than narrow. The 2026 AI Index put the estimated annual consumer surplus in the United States from generative AI at 172 billion dollars by early 2026, against 112 billion dollars a year earlier, with the median value per user tripling over the same period. Adoption figures are contested and depend heavily on how the question is asked, but the pattern is not one of a technology available only to a few. A reader who thinks concentration is the central risk has to explain why the benefits have spread so quickly while the control has not.^[4]

Finally, the standard remedy has its own concentration problem, and its advocates say so. Sastry, Heim, Belfield and their co-authors, who make the case for governing AI through compute, write in the same paper that naive or poorly scoped approaches to compute governance carry significant risks in privacy, economic impacts and the centralisation of power. There is no version of this policy in which the power to allocate compute is widely held.^[9]

The European Union's AI Act imposes obligations on general purpose model providers, and the prohibitions in its Article 5 have applied since 2 February 2025. The timetable has since been loosened. Council presidency and Parliament negotiators reached political agreement on a digital omnibus on 7 May 2026, moving the obligations for standalone high risk systems listed in Annex III from 2 August 2026 to 2 December 2027, and setting 2 August 2028 for systems embedded in regulated products. The Council gave final approval on 29 June 2026.^[16]

Competition authorities have documented rather than intervened. The Federal Trade Commission's January 2025 staff report is the fullest public account of how cloud providers and model developers are tied together, but it was a market study under Section 6(b) rather than an enforcement action, and it was published as the agency changed leadership. Nothing in the record so far has unwound any of the arrangements it described.^[6]

Corporate structure has been the other live venue. OpenAI completed its recapitalisation on 28 October 2025, converting the operating company into a public benefit corporation called OpenAI Group PBC controlled by the OpenAI Foundation, which holds about 26 percent of the equity, worth roughly 130 billion dollars, with Microsoft holding about 27 percent. Reporting by CalMatters set out the objections: the same people can sit on both the nonprofit and for-profit boards, and the agreement with the California Attorney General is short on detail about how the foundation stays independent. Steven Adler, a former OpenAI safety team lead quoted in that reporting, said he hoped a truly independent body would do a better job of protecting the mission than one that feels any pull toward profits.^[17][18]

On the research side, Davidson, Finnveden and Hadshar propose concrete measures: rules in model specifications and government contracts that forbid assisting a seizure of power, information security against unauthorised access and the insertion of secret loyalties, sharing powerful capabilities with several independent stakeholders rather than concentrating them, procuring military AI from more than one provider, and distributed control so that no individual commands enough military AI to act alone. The International AI Safety Report 2026, led by Yoshua Bengio with 91 co-authors and an expert advisory panel nominated by 29 countries plus the United Nations, the OECD and the European Union, reviews concentration of power among the systemic risks it covers.^[1][19]

Keep the measured and the argued apart. Investment concentration, cloud market share, chip supply and political spending are all counted and public, and they point one way. The claim that these produce a political settlement that cannot be reversed is a projection built on assumptions about how far capability will substitute for people, and those assumptions have not been tested. A reader can accept the first set of facts and still think the second step is unproven.

The most informative indicators are institutional rather than technical. Watch whether legislatures retain the practical ability to write rules that bind the largest developers, whether any government can obtain independent evaluation of the systems it buys, and whether the gap between open and closed models widens or narrows. Watch also for the reverse case: a serious antitrust action, a state law that survives federal challenge, or a procurement rule that requires more than one supplier would each be evidence that the ordinary checks still work.

Note who is making the argument. It is unusual for the chief executive of a leading developer to name concentration of power as a top risk and to include his own firm in the category, as Amodei did in January 2026. That does not make the claim true, and self-interested reasons to make it can be constructed, but it does mean the concern cannot be dismissed as the property of one political camp.^[3]